Machine Learning Approaches to Predict Type of Rehabilitation Based on Pre-admission Clinical and Patient-based Outcome Measures for Spine Surgery

Introduction: We aimed to develop a machine learning algorithm based on pre-admission variables that can accurately predict the type of rehabilitation in patients undergoing spinal fusion surgery.

Methods: The Stanford University Hospital database was used to select patients who underwent a spinal fusion between 2014 and 2022. Our primary outcome measure was the type of rehabilitation facility: home-based v/s non-home-based (non-home-based rehabilitation included all facilities, including inpatient rehabilitation facilities, skilled nursing facilities, and unskilled nursing facilities). Five machine learning algorithms were developed to predict rehabilitation type and were assessed by accuracy, discrimination, and overall performance.

Results: Two thousand nine hundred and eighty-six patients were included. The median age was 67 (interquartile range [IQR] 58–73), and 46.8% (n=1400) were female. The non-home discharge rate was 31.2%. The percentage of patients whose procedure was categorized as “elective” was 96.0%. Our models included age, sex, race, BMI, procedure, number of levels of fusion, spine segment, admission type (“elective” or “urgent”), and operating surgeon as variables. XGBoost was considered the best model based on discrimination (AUC = 0.70), and overall model performance (Brier score=0.224), while other models had the following performance: Random Forest [(AUC=0.69, Brier score=0.234)], Logistic regression [(AUC=0.68, Brier score=0.234)], Neural Network [(AUC=0.65, Brier score=0.271)], and Decision Tree [(AUC=0.65, Brier score=0.299)]. Through Shapley analysis performed on the XGBoost model, age at admission, number of levels of fusion, and gender were the variables that contributed the most to the predictive value of the model.

Conclusion : This preliminary result has shown that it is possible to create a predictive machine learning algorithm with good accuracy to predict rehabilitation type based on pre-admission variables. Using our methodology, this model can be further developed for other conditions and treatments.